Skip to content
Bud's Wiki

Project Eden — persistent multiplayer world model that decouples world state from visual rendering (Tripo / VAST AI Research)

TL;DR

Section titled “TL;DR”

Tripo (powered by VAST AI Research) announced Project Eden, a research-preview world model that is pitched as a persistent, multiplayer environment whose underlying world state is decoupled from visual rendering. Instead of treating a world as a sequence of transient frames, Eden treats it as a structured evolving environment that runs continuously, can be modified by user actions, and can be consistently observed from any viewpoint. The tweet is a product-style announcement with a demo video — no architecture, parameter counts, training data, or benchmarks are disclosed. It is the third filed datapoint in 2026 of a multi-player WFM design that hoists shared state out of the network, alongside Odyssey’s Agora-1 and the MultiGen research paper.

Key claims

Section titled “Key claims”
  • Project Eden is a persistent world model that “runs continuously” — state evolves across time rather than being regenerated from prompts per query [tweet body].
  • Project Eden is multiplayer: multiple users can act in and observe the same shared world [tweet body].
  • Eden “fundamentally breaks from existing paradigms by decoupling the underlying world state from visual rendering” — the world is treated as a structured evolving environment rather than a sequence of transient frames [tweet body].
  • The decoupled state can be modified by user actions and consistently observed from any viewpoint [tweet body].
  • No architecture, model size, training data, training compute, latency, framerate, resolution, or benchmark numbers are disclosed; the only released artifact is a demo video and a research-preview framing [tweet body].
  • Authorship is attributed to VAST AI Research (the team behind Tripo) — Tripo’s stated focus is 3D + AI, so Eden extends that lab’s track record from 3D asset generation to interactive worlds [tweet body; @tripoai bio].

Method

Section titled “Method”

The tweet describes Project Eden only at the conceptual level: a world state component that runs continuously and is updated by user actions, plus a rendering component that produces per-viewpoint observations from that state. No model architecture, training objective, or system diagram is shared in the announcement. Multiplayer behaviour is described as multiple users observing the same world from different viewpoints rather than as a specific architecture (e.g., shared cross-attention, per-player rollouts, etc.). The viewpoint-consistency claim — “consistently observed from any viewpoint” — implies that the rendering pass is conditioned on the same authoritative state object across players, but the announcement does not say so explicitly.

Results

Section titled “Results”

None disclosed. The published artifact is a teaser video and a research-preview announcement. The tweet reports 3M views, 597 likes, and 160 retweets at filing time, which is signal of community interest but not of technical performance.

Why it’s interesting

Section titled “Why it’s interesting”

Project Eden is the third filed 2026 datapoint in a converging architectural pattern: hoist shared persistent world state out of the neural network and condition a learned renderer on it, rather than absorbing all multi-player interaction into one frame-sequence model. It joins Agora-1: The Multi-Agent World Model (Odyssey’s closed-flagship 4-player GoldenEye preview) and MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines (MultiGen’s open-research diffusion game engine with explicit memory) on the same design point, in contrast to Solaris: Building a Multiplayer Video World Model in Minecraft‘s single-AR-DiT-with-concatenated-participants approach. With Tripo / VAST AI Research entering this design space, the “decouple simulation from rendering” pattern is now present across an open research paper, a closed AI-video lab (Odyssey), and a 3D-focused lab (VAST) — three independent provenance points, which is what BUD.md §6 calls a pattern rather than a coincidence. It also extends the World Foundation Models cluster’s product-surface taxonomy: alongside Project Genie (single-player closed flagship) and Waymo’s vertical specialization, Eden is the first filed datapoint from a 3D-asset-focused lab (Tripo / VAST) reaching into interactive WFMs — a different lineage than the video-generation-first labs (Runway, Odyssey, DeepMind).

See also

Section titled “See also”